The AI Revolution Call

Foreign.

Welcome to the AI revolution called where we also discuss robotics in space from time to time, but mostly AI. I'm Bryce Smith with 10,000 Days Capital Trading with Cody. And with me today is John de Olvera. Hello. Aix. Our resident AI expert veteran. And Craig Galant. And so welcome guys. Welcome back.

Thank you very much.

It's Christmas this week and so all of my family members are getting AI gifts somewhere. Probably chat. Chat GPT Pro subscriptions. I don't know. Nvidia GPUs.

That'd be great. So you have one in the mail on the way?

Yeah, yeah. I got an H100. You know, just drop, drop shipped it to my, to my cousin's house.

You know, if you can find some memory, let me know.

I kind of met on the way to me but that's okay.

Yeah. John, how are you? How are things? Any. Any updates on what you're working on or what's got you excited right now?

I've been excited about, about the new, you know, Google graphics capability kind of thing. So let's see. Wanted to show you. I did. This is one shot. Can I share this? Let's see.

It's.

It says it's sharing.

So let's see. Right.

Yeah. There you go. You see it? So I put in. I actually used Claude to make the. I gave it a short prompt and I, I said make a better prompt out of my prompt, you know.

Yeah.

And then I gave that prompt to Google and said to, to Nano Banana Pro. Just going into, you know, AI Studio, I said make me an infographic that describes the process that our development team is trying to use wherein it's hybrid org structure of humans and AIs and yet the humans are accountable for it always. So at any stage of output you can say why is this like this? And you know who to turn to and who's responsible for it being that way. So you see a workflow along the kind of development workflow along the left you see an org chart and upper middle. And then you see the different kinds of engineers working and how they're sharing information there. I just can't believe that this is like one prompt. I say one thing and it goes and it makes this.

It's truly amazing. I mean, just was it six months ago that the AI couldn't even make pictures that had legible words?

Right. That's a huge aspect. That's a huge piece of the breakthrough is words. Yeah.

And it would always have this like alien text and stuff. And it was just like a known problem and everybody's like, yeah, it'll, you know, someday it won't have this problem. But now it's like that's again seems like the dial up Internet compared.

Look where we are. Look at the change that happened in those. This is why I feel like this is like this, the chat GPT moment for images. They're suddenly so useful you can do anything.

Do you have any idea, John, technically what they changed to make it do.

At a high level? I know one thing which is before this stuff was purely using diffusion models where it's starting with poltergeist snow on the screen and gradually moving that in the direction of a picture, almost like running a compression algorithm backwards. So when you compress stuff and compress it until it doesn't even make sense anymore, you end up with the snow or the noise. And then when you say I want to do that kind of thing, like what would this look like if it were such and such. And then it starts to piece it together. So that was the diffusion algorithm and that got us so far as the original OpenAI image generation and some of the other. I can't even remember the names of them because it was at least over six months ago to see if these things go.

But you remember what's that ancient history at this point.

Yeah, there were the ones though they, you know, people using from within Photoshop and different things. But so any that's got you there, what got you here was that being combined with like Gemini 3 to where the regular type LLM is describing what it needs for each piece of a thing together with that diffusion stuff happening. So that combination lets it do this and also lets it be so smart that you can on another pass tell it that light bulb in the upper left, make it red instead of yellow or whatever and have it change just that and not mess. Because right now too there's the problem where, you know, you change something to try to fix something and it just messes everything else up so that it doesn't do that and it's cleanly changing what you want. And also to where you can say like here's a dozen features of an image, this person, this shoe, this dress, this whatever, and say make a picture that incorporates all that stuff and, and it'll do it, you know, or here's a person like sitting in a room, show me that person from 10 different angles, you know, above and behind and around and whatever and. And it's the same person. It doesn't like mess it up. It's like character consistency called Right. So all of that exists now. And also you can take the, you know, that, that consistency and use it to make different scenes. Like with open AI. It's called Sora, the. The video model with Google, I think it's called VO. Yeah, VO. And you can say, like, go make this 10 second clip or something, and this person's going to run from here, over there, so forth, and then grab the last clip and use it as this image starting point for the next set of stuff that you do. And just get, you get the kind of the continuous consistency through video as well now. It's just. It's just crazy.

Yeah. You know, I've seen a good amount of AI videos throughout the year as this has progressed across, like what you just said from Sora to VO and everything. And it's. I can't remember who. I was talking to someone over the holiday at Thanksgiving and I was like, you know, I count myself as like savvy young tech guy and I can tell the difference between AI and. And the real video. But there's been a couple videos in the last couple weeks where I really had to like, go to the profile, kind of scan through it and see like other videos they've made. I'm like, oh, no, this is AI. But we've crossed. I think we can, like, if we haven't crossed it yet, within the next six months, we cross the frontier of at least in short form, you know, 15 second clips that are indistinguishable, you know, from AI. Yeah.

What is. It was called the Turing Test, you know. Yeah, it'll be like the visual Turing Test.

Yeah, exactly. You know, it used to, like, they would have like six or seven fingers or maybe like their arm would be a little bit too long if you're looking at a human or something, you know. And now it's just, you can't tell. I mean, it's like, like I think I saw was like a skiing video or something like that. And it was like, I can't remember. I already said this on the call. I don't remember, but it was like, you know, skiing is like, like a GoPro version of a skiing video. I mean, there's not that much to it in real life anyways, you know, and so for AI to make that, it was like, it was, it was truly indistinguishable from real life. I.

It's hitting the point where I understand people who get annoyed at it because I was annoyed when I see, you know, like a good pet video, you know, like the Animals being funny or whatever. And so I saw one yesterday. This, you know, chicken walks into a living room where these four dogs are sleeping, and there's a lady sitting on a couch. And I was fooled because the chicken walks in and carrying a balloon. Let's go. And like, it's possible it could happen. And so. And the moment the balloon is like, pops or is. Flies over something, the dogs jump up, they knock over the table in the middle of the living room. The lady starts screaming, the stuff's doing whatever, whatever. And then so I'm like, got to check the comments. I go to the comment, they're like, AI, AI. And they're saying, like, look at the woman's mouth. Gets distorted at this point.

Whatever.

So you look and you're like, yeah, that's fake.

But.

But it's like really realistic.

Yeah, yeah, I saw that same video and I thought the same thing because I was like, there's no way that it got that crazy that quickly. But I tell you what, if I saw. I saw two videos in a row of like, random dogs saving a baby in a house fire, and I was like, what's up with all these dogs saving babies? And I realized, I'm like, yeah, AI.

But it's. That's how good it's gotten though, because it really fools you, you know, and so it's a little. It's a little unfortunate, you know, I want the stuff to get better and better, but I really want to know when, you know, when it's. Which I don't like, because it matters, you know, like, did this happen in the world? Is.

Is.

It's not as funny if it wasn't real.

Right?

Do you think it's even possible to, like, let's just say, hypothetically, the government's like, we have to have a disclaimer on all AI videos. Is that even possible to.

Yeah, I mean, this stuff is watermarked really well. So like Google, at least they did stuff with their images where you. If you could crop it and resize it and cut out a piece and send it through another thing or whatever it will, they will still know that it was produced by Google's image generator because it's got stuff buried in it. Every pixel is a number between 1 and 16 million or something. Every individual pixel. You can shift pixels here and there and make them add up to something and hash out to something, and you would never know or find it. I think they have patents around it even.

Yeah, I think one of the interesting things and one of the intersections of AI crypto. I can't remember if we've talked about this on this call or not, but potentially maybe it was Mark Andreessen that said this, that you're going to have blockchain tied to all of the image and video somehow so that you basically have a, you know, source of truth as to whether this was real or not. And if it doesn't have the blockchain tied to it, then it's, you can't verify its, you know, authenticity. I don't know how it would work. I mean it would have to be built in natively to like the phone, I guess, or the, you know, whatever the actual image creation creation was. But I mean, in theory I understand where he's going with that.

I mean a blockchain address is just a number and a number can be encoded like I'm saying, like across pixels and be redundant and have to, and be, you know, made out of a ratio. So it's not about whether it's even that exact number. You know, different things. And then I think they could, I.

Think they could do it. So to the, to the person, like to the average Joe, they still not be, might not be able to tell. But if it went into like a forensic, you know, type lab or something like that where somebody was, you know, cross sectioning it trying to figure out, they could definitely figure it out.

Or there's a, an API call, there's an app like, associated with like send this image here and we'll tell you this about it kind of thing. So.

A little bit different topic but John, one of the things that we're working on is like like a top five predictions article for 2025 and Andreas and Horse woods actually did this format and I really liked it. But they had like everybody on their team was like, you know, give us your top five things that you're kind of expecting with AI this year and it was just, or yeah, in 2026. So I don't know if you've thought about that much, but kind of I, I, we've asked you this question before just, but just over the next 12 months. Where do you, you know, what are kind of some of your predictions? I know this is off the cuff and maybe haven't thought about this, but what, what are you expecting to see? I mean, I know, I think we asked you this last year and I, one of the things specifically I remember you talking about was multimodality and that absolutely did happen almost, I mean every one of the frontier models now is multi modal, natively it can do image, video, or maybe not video, but it can do image, text, voice. And so that's sort of like, yeah, it's table stakes now. To be a frontier model is multimodality. So we're what, you know, if you can tell us maybe what are a couple of things you think you expect to see over the next 6, 812 months.

Well, I mean, you know, I come at it from a particular angle. I'm not, you know, a finance person. So as AI developer, which is kind of, you know, where my focus is, what I'm doing each day, agents are the big thing, they've been the big thing, they'll continue to be, but they're getting, it's getting better at being predictable with its tool selection and outputs and so forth. So that you can do, you can get something to do something for hours and do a good job with it and come back and you know, so that's something I see happening. I see the beginning of this AI based engineering methodology. Instead of, it's like taking this notion of vibe coding where you put in a prompt, you say a little something what you want, it goes and makes a program. But instead that using AI with an engineering team. And it was like in that image I was showing the whole team working with AI in different ways to do all of the aspects of the software development life cycle. And that means speeding everything up, getting expert software engineers to be contributing at what they're good at, getting junior developers to be contributing what they're good at. Product managers to be able to take things from like we're working on a piece of our product that got taken from idea to project plan to stories, which are the tasks for the programmers, like all the way through all of those with AI assistance at every stage. So I see that being a huge piece is, is that getting going. I think I see Anthropic doing well because they've been kind of cornering the market with the best tools, combination of tools for developers. I see Google continuing to do really well. I was the, the Google pessimist in a way. Not because I didn't think they're, you know, smart, but I just think that they tend to step on their own feet when it comes to product launches. So, so, but I see them doing, doing really, really well. I don't know about OpenAI, you know.

Really?

Yeah, I don't know. We'll, we'll see what happens. They, they, the model they came out with that they just came out with 5.2 is really in particular aimed at the knowledge worker, the white collar person sitting behind a desk and taking some of those jobs. You know, I see more, more job loss in 2026. I see it, I see job loss before job gain. You know, there'll be plenty of new jobs and it's just a question of like when that kicks in.

Interesting. On the first thing you said with respect to agents and a lot of people said 2025 would be the year of agents. So what did that? Is that. Are we just a year late or was 2025 the year of agents and it's just going to keep going, accelerating?

Yeah, I think Karpathi or somebody was like agents is going to be from now till 2030 or some number of years. And I think it is a multi year thing. And agents did happen in 25, but they sort of got going and it was more individual or several agents, but it wasn't. But now it's becoming, you know, an agent can handle a hundred or hundreds of, of different skills and tools that it makes use of. And then with the agents, how they're talking to each other, it's becoming like an operating system of, of agents. So I think it's just a sophistication level. It's, it's just leveling up. It's been happening in 25, it'll happen in 26 even more.

What in terms of how these agents work is it. Do you have one agent like the agent the user interfaces with? Does that agent go out and accomplish all of these tasks? Or is there almost like a hierarchy and they have sub agents that they can call upon to do specific. Say if you're a software developer in an enterprise and you want to get something done, you tell this agent and then it calls on maybe the two or three among 12 agents that are the most relevant to accomplish the task. Like how does that technically work?

Yeah, it, it is a hierarchy. We typically, and with us it used to be more of a hierarchy, now it's more of a flat hierarchy. So there's the agent that you talk to and then it calls upon a set of other agents, but not levels on down. Like they don't then call agents which call agents. It just calls to those other agents and those agents. That's our architecture. And I see more of that rather than it going, you know, nested very, very deep. It's. There's been some move that I've seen. There's been some move away from that.

From the, the nesting multiple.

The nesting multiple because you can like you can take an agent and it can handle so many tools that you lose kind of the thread if you keep handing along and it doesn't know what to go do with that, etc. That there's what happens in an app that does the AI stuff and then there's what you might as a developer, like we use claude code where you're, you know, on the spot making agents and making them go do things as opposed to. Right. So the first kind you're actually coding agents into your application and the other kind you're doing development work and you're saying, hey, go launch a sub agent to blah blah blah, and just kind of making it happen right there.

I mean, that makes sense. I think from an enterprise standpoint, if you have to build out like this very detailed hierarchy of agents and tell it how to call one to do something in Salesforce and something in another platform that just gets to be too laborious and cumbersome for an enterprise to actually implement and be useful. But if it is more flat like you're saying, I would just think as the models themselves become smarter, they'll be able to accomplish more tasks on their own and delegate better the end user. An employee at a big corporation can just have the AI actually do things without having the IT department program this very elaborate scheme of, of agents. It seems like that, I mean, I feel like that's probably one of the thing that's held back enterprise adoption to some extent.

There's also the kind of sophistication that's gone from prompt engineering, as they called it, where you're, you know, you give it a prompt, tell it something to do and it goes and does it, which is what we're all just doing whenever we're chatgpt as well. From prop engineering to context engineering, where there's more careful selection of what all information is being provided. I attached this file, I gave you this data, etc. And now go do stuff with it so that there's that kind of thing. And now we're moving from context engineering to something about. I've heard the name involving harness. It might be like harness engineering, but it has to do with what's that whole setup. What's, what's kind of the architecture of this talks to that and gets this done, which puts the data over there. So it's less about what model am I calling and more about what's the whole setup.

Have you seen any kind of like, I mean you're, you're at an AI native company and you do this for a living, but Is there any pushback within companies? Like people don't want these tools, they don't find them helpful, they don't find them useful, they're afraid of them. I mean, I know some of the people I've talked to. One kid I talked to as probably six, eight months ago was like, oh, I don't want to use AI. He worked at a big consulting firm. Because then he was like, well, I will just, they'll figure out that AI can do everything I do and I'll lose my job. And I was like, man, I think that's completely the wrong approach to take. But what, you know, what is your like the end user? What has kind of been the response from them?

Well, what's interesting is yes, some end users but also some software engineers that are, have, I mean, I derisively say Luddites but have a certain attitude toward it. And it's not even that, oh, it's going to take my job because it can do everything. It's more that like, oh, those dumb managers who are going to let this thing come in and willy nilly create code that no one understands what it does and that kind of thing. So it's kind of rejecting the positive aspects because it's true. Yeah, it can write bad code, but it can also write good code. And if you're doing a good job with it, you're kind of getting it, you're seeing where it does which and helping it do more of the good stuff. But happens. That's funny thing is it happens even with the technology people.

Yeah, it's interesting. I know a guy, he works at a data center here. It's a government data center that does a lot of the processing for the National Reconnaissance Office, which is the division of the government that handles the spy satellites. So he can't tell me too much about what they do in specifics. But like 8 months ago he was like that guy, very reluctant. Um, did not like, he was kind of, you could just, was very terse when we talked about AI and you know, was kind of a naysayer and I was speaking with him at a holiday party this week or last week and now he's like all in. He uses Claude. He's, he's, he does a lot of like software testing, I guess. And so he has to test what, whatever they've made and see how it performs, you know. And like he says, he uses Claude every day. It designs the frameworks to do the testing. He's building apps with replit. It was a complete shift in his attitude towards AI, and he's really come around to it. I feel like probably a good part of that is he saw the AI improve and it was useful. And it wasn't like what you said, where it's, oh, it's building thousands of lines of code that nobody can understand or know how it works. And he's like, no, actually, this does make sense. This is making me more efficient, and it is helping speed up the testing process of software development. So, I mean, testing is one of those nodes or whatever you have in your graph or your picture you showed at the beginning, Right. I mean, before you push software to production, you basically have to do a lot of testing to make sure that it works like you expect it to, right?

Yeah, there's a lot of work involved in all of the steps of software engineering and development and deployment and all. Whatever. There's a lot of steps. A lot involved in each of those steps. And so if you get. If you can get help with creating a test suite, creating the code that runs over that test suite, doing all the different pieces of things, it's not just, here, go make code for me. It's just that all these aspects of IT to be involved, and it can help in so many ways.

Yeah, and I. I understand the sentiment. If you're a real user or an IT professional, like, you don't, like, imagine if Excel formulas, like, were only accurate, like, 80% of the time, that's great. But I can't rely on Excel formulas if they don't do the job 20% of the time. Right. And I feel like for him, finally, he found the AI had reached that threshold where it was dependable enough that he could use it consistently and actually benefit him and his job, because, especially as a government employee, I mean, your incentive is to make life easy. And not necessarily. They don't need to adopt the cutting edge if it's not actually helping them in some way or another. But he actually is finding real value coming from Claude specifically now for what he does. So I thought that was an interesting anecdote.

I wanted to ask John. So we had a really good conversation in our Discord channel last week. It was two guys, and they were talking about how, you know, one of them's like, man, I just, you know, AI, it just always gets things wrong and I just can't get it to work. And then the other guy was like, yeah, but, I mean, I just work with my AI, I tell it when it's wrong. It's almost like raising a child. And then over time, it starts to understand what you mean and it gets better and it learns what you like and all this other stuff. Do you find that with coding using Quad, it's similar in that aspect.

That's why I think where context engineering came from, that notion of it's really not about, you know, the model being smarter. You get to a certain point where the real thing that affects how good and useful your output is going to be is your input. And when your input was just, you know, write me a poem and you're not telling it anything about what you want to actually take place, then you get what you. You get out what you put in. But when what you put in is really like real structured data and specific instructions and so forth, you'd be amazed. It keeps up with you no matter what you go and do. And the more you do that, you're like, wow, it actually knew what I meant and did something useful with it kind of thing.

Yeah, that is so true because I've used, We've been using replit to create some app and I'll go to Gemini to, like you said, have it help me create a better prompt that can kind of go over all these other pitfalls and then give it to repl it and it just comes out swimmingly. So, I mean, it really is just about making sure the input is perfect.

So, like, this stuff that we're doing and talking about, what percentage of people are doing that one? Like, I think, you know, the majority of people, if they've gotten on at all, are doing the typical, like, they're, they're googling with chat gbt and you don't have to be a programmer. A lot of folks now in the vibe coding world aren't programmers. They're, you know, people like entrepreneurs, you know, marketing, digital marketing people. They're like, they're just like, how do I find an edge? How do I go and do something that kind of advances me by, by using this stuff, those people are taking advantage of it, but it's just a very small percentage. And we need to see more of people on the leading edge doing more and then more others seeing what those people are getting out of it and then them getting in and stepping beyond what they're currently doing. Because I just feel like, I think, and I think somebody was saying, maybe it was all Sam Altman was saying, you know, like, whatever people go and do, we get the new model out there and so forth. They haven't unlocked. There's 90 of it. They have not unlocked yet. Is so Much more it can do than people have done with it.

It's sort of like the classic thing like even with the iPhone, it was like Steve Jobs has said, people didn't know they needed a smartphone until we gave it to them. And the capability of the smartphone led the demand for it by a long shot and huge margin. And eventually the app revolution took off and people started thinking of all these cool things they could do with iPhones. And AI is probably going to follow that trajectory in some way. Some ways, like the models are just going to keep getting smarter and more capable and people inevitably will figure out more things to do with them and use them more and more frequently accordingly. I think it would follow that same path.

I picture kind of a snowball rolling down a hill and the more things it can do, the more apps, the more features, functionality keeps coming along and that's kind of growing the snowball and then the time passing is it going down the hill and that's growing the snowball. So it's kind of growing it from multi, multiple like dimensions. And all of a sudden you've got this really fast, really big thing like hitting everybody.

Yeah, absolutely. Do you think going, circling back to your, your kind of your predictions and the anthropic versus OpenAI versus Google thing, I mean that, that is talked about a lot in the markets because OpenAI has signed trillion dollar plus worth of compute commitments to Oracle and Core Weave and others and people are questioning whether they'll actually be able to meet pay those commitments if there's actually going to be demand. I mean to me it seems like it just really feels like putting the cart before the horse in terms of like open AI in order to grow needs the compute. And so is it really like I don't think they're being, it's not like they sign these compute deals like for headline just to generate the headlines. They actually, if the growth keeps up and what they expect and reasoning models keep growing and tool usage keeps growing and the amount of tokens, you know, you just plot that out. Like I, I understand why OpenAI would want to contract with all of these companies to try to secure compute. It's not, I don't think there's anything like malicious or stupid or something that they've done by signing all these compute deals. Maybe, maybe it's going to turn out, you know, could cause an overbuild in the amount of data centers that end up getting built. But in regardless of whether you look at OpenAI or Amazon or Microsoft or anybody, there's A shortage of compute. And so racing to get as much compute as you can seems to actually be the logical step for them to take at this point in time.

Yeah, I think we've said that it seems like there is as much demand as there is supply. And as the supply goes up the demand will, will stay with it. And so they have a much bigger worry about the supply side of it. Can we even get enough electricity and data centers online to keep up with what the demand could be? Because they need the demand to be high. They need, they need to fulfill that high amount of demand in order to keep pace with the expectations.

And then I think too the second part of it is just I keep coming back to Nvidia over and over again just because I, I say that Nvidia is the AI revolution, their AI revolution is Nvidia. Maybe I should add to that Google and Nvidia are sort of the two because if it weren't for the extraordinary, you know, gains in compute that came from the Nvidia parallel processing architecture, none of this would would have worked. The transformer would, it would have been meaningless if you couldn't, if they didn't have GPUs to run Transformer architecture. And for from OpenAI's perspective having access to the Nvidia compute first and having it early, having large enough clusters to train the next order of magnitude, two orders of magnitude models. If you're talking parameters, maybe it's a trillion parameter model or something like that existential for them. I think like if you, if, if they fall behind in that race then their models lag and they're going to use lose users. I feel like.

You do see like Google's own processor starting to get talked about more and more. Sure. You know the, the tensor processing unit is there. Does it matter as much you know, if you're Nvidia at that point, if you're using Nvidia?

I think so. Just because the main reason that the TPU came about is because Google wanted a processor that could hand like specifically optimize for their use case, which is not necessarily the same as what Nvidia optimized for. You know, I mean search and the machine learning that goes into search is not the same as powering a large language model. And I think TPUs are very capable and Google has been done outstanding. I mean you know we're huge Google shareholders, it's our second largest position. But in terms of just building pure AI, I think Nvidia and the best AI, it sure seems to me that Nvidia still has that edge. And if you're renting compute, they absolutely have an edge because other developers know how to use Nvidia GPUs. They don't necessarily know how to use TPUs. Yeah, I mean that's, that could change. But if it wasn't for the shortage in the GPU market for now three years, I don't know that anybody would be, you know, that TPUs would even be talked about. It's just the mere fact that you can't get access to GPUs like you want that people are looking for alternatives. I mean, plus the fact that they're expensive, there's no doubt about that, goes.

Back to that supply kind of idea. You know, they've been, if there's been a shortage for years and then other stuff starts popping up because of it. And Nvidia is like, hey, we just need to do everything we can to keep pace on, on, on the Demand. It's like, it's like a waiting list who, who, who gets to have a product where they're, you know, the on the demand side is, is like a waiting list of people wanting it.

Right.

It's like one of those expensive restaurants that, you know, it takes, you know, a year to get into because they're booked that far in advance kind of thing.

Right. Well, while we were talking, I had Chat GBT with Replay. I don't know if you've, if you've seen, you now can call Replit within Chat GBT and make an app. So I had Replit make a Spotify wrapped version of AI in 2025. So I'm just going to show you guys this. I don't know how accurate this is, but I thought it was, it would be a cool idea. Let's see. So this is your year in AI exponential. I told it to use real data, so let's hope, hopefully it uses real data. But 14, 200 new models on hugging face this year, I could believe that.

Yeah, that's believable.

And 2.1 global users probably at least. I mean chat has 800 million users alone. And then Google has, you know, I don't know, 2 billion people use Google every year. 3 billion. And Google is using AI through and through. From AI overviews, AI mode, Gemini, everything. Let's see what's next. Model breakthroughs, average context will windows 2 million tokens. I don't know that that's true. I know Google Gemini is 1 million token context window.

I believe there's, there's one out there. With 2 million.

But the average, I wonder.

Oh, average. That seems high.

Yeah, that seems high. Reasoning costs. Absolutely. The price per token has come down significantly this year and that was the deep seat moment was everyone was worried that, oh, they built this model really cheaply, it's going to bust the AI thing. And actually everyone in AI was like, no, we want reasoning costs and token costs to come down. Right. GPT5. I do not think that came out in 2020. February. I think that was later.

Well, there's been 5.1 and 5.2.

Okay, fair enough.

So.

Inference became cheaper, faster and more ubiquitous. H200 and Blackwell chips are the standard. Yeah. 45 efficiency improvement in tokens per watt. Probably that's an understatement. I mean I think Blackwell is like three or four times more compute power efficient per compute per token. Training flops 10 to the 26. I don't even know what number that is.

It's very, very big. Yeah, so. And it's actually hard to believe.

88% of Fortune 500 companies have AI in production. Maybe. I mean everyone in corporate America is using AI like chat GBT and stuff's chat bots. I don't know if they have like custom apps that are at 80.

Yeah, but they would count that. They'd count the chatbot stuff.

Right. 65 of all new code written today is generated by AI. What are your thoughts on that?

It's too high.

Too high.

I don't believe it. I mean, I believe we're headed that way, but. Yeah, yeah. Just where's this? Does it have any sources for what it's saying?

I don't know. Yeah, open source. Llama4mistral proved open weights can rival proprietary giants. I actually think open source has been sort of in a bear market. I feel like Llama has declined. Like, I think Llama came into the year as like one of the leading models and has sort of lost its luster.

What about like a deep seek or a mixture? I mean there was still a lot of open source, but yeah, I think it maybe hasn't met the very high bar that people were starting to set for it.

Yeah, Meta themselves has sort of like backed off the open source. Yeah, like champion. And they, they. I think Mark Zuckerberg said we're always going to be like there's a difference between open weight and open source. Like open weight meaning you give away the weights but not necessarily all of the data and fine tuning and everything that went into it.

Yes, Right.

Regulation and safety. Global AI treaty signed in Geneva. I Did not know that if that's true. Watermarks. We were talking about that AI and culture first. AI generated chopping songs. That is true.

That's. That happened.

Yeah. I don't know that 15% of top 100 hits are AI generated but there's certainly been a few.

The first blockbuster movie with. Well, it's just talking about the backgrounds. I would believe that.

Okay. Again, this is going to continue to accelerate. 2026 prediction AGI 45% this year. Next year seems optimistic.

That's 45% of people believe that 26 will be the year of AGI.

I guess so yeah.

People just want to see it here.

That's a wrap.

That's a wrap.

Data source there though.

Agents become autonomous employees.

Like basically what you were saying earlier.

I mean it's a big difference between saying I just ran this piece of code and it went off and did something versus I've got six people on my team, three humans and three agents, you know.

Yeah, but this is pretty cool. I mean again at the beginning of the year if AI couldn't even make this Spotify wrapped thing that I asked to make on replit via chat gbt. Like I remember, I remember when I wrote the article about Google one like I was able to make a Snake game in one prompt and I was like super impressed by that. I was like, wow, I've tried AI. I've tried to get AI make a Snake game before with one shot and it's never been able to do it and now it can and we've moved so far past that and that was in March or April. Now AI is actually like. I mean video game developers are using AI all the time. People are making. I mean what next year probably we get commercial video games that are AI generated and.

The, the, the role of worlds will keep increasing. The notion of kind of creating a. It's like virtual reality stuff but you're making a. A place that can be viewed from any angle, that can be walked through and then psych. Those, those video games that are typically have you exploring some space where you know you're first person shooter kind of angle on things and you go along and go through here and up this hallway and down those things took a long time to design and build and all the artists doing their pieces together, that stuff is becoming automatic. You say I want this and you go start walking through it and it's making more of it as you go.

Right. So cool.

Yeah, I, I definitely think that's where gaming is. I think next year. So we Already have a few games out right now that you can, can get download on Steam that have a lot of AI generated content and also leverage AI for storytelling. But I think next year is the year where it becomes almost ubiquitous because there's a lot of pushback in the gaming market. Just the, I don't know if it's your average gamer but the loud voices on X and stuff, they're like no AI in the games. And, and you know, I'm just sorry to say that's just where things are going, you know, especially when you're in such a, a big R D or kind of expense heavy thing. You know, it's, it's, it's hit driven. You know, when you make a game you're trying to make a hit and that's where you make your money. So they're going to use every tool they can in order to make sure those, you know, their costs are down.

The thing with them saying like no AI whatever. The other place I hear that kind of torches and pitchforks kind of thing happening is when creatives are using AI. When you see companies come out with commercials that are built using AI and then all of a sudden people are like you're putting graphic artists out of work. You're putting such and such like you hear it a lot from that side of things. You know, Coca Cola does a commercial and Budweiser does commercial. And like one of them it works because they weren't like emphasizing it in a particular way, relying on, but it was AI. Another one, it's like what did you do? This used to be a good product. I'm not buying you anymore because you went and did this thing or it's a lot of anti AI sentiment in areas where you can clearly see the connection to job loss.

Right. Well, it's interesting I've been saying that AI has the potential to be very egalitarian and level the playing field. And I just saw it today. I didn't watch the actual advertisement but like a small bakery or somebody is using AI to make these like car chase videos where like this guy's like racing from the cops and it's like a high speed news or high speed chase and the news is reporting it, but he's just racing to get to the bakery to get the donuts or whatever before they're sold out. That bakery, random bakery in Florida or whatever it is, would never be hiring a graphic design team team to make a Coca Cola style ad.

Yeah. So who are they replacing? Yeah, who are they hurting? They, they're just doing something that wouldn't even have been done.

Right. And that's sort of, that's sort of been the case with social media with you know, a lot of technology. Is it just there's a long tail of indicate into users that can now access this technology because it's so cheap and ubiquitous.

You know, the whole influencer industry. Right, Yeah, I mean that, that's all based off social media. And like you're saying, and who's to even say that bakery would even think about doing a commercial. They were empowered because the technology is literally thrown at them to do it.

You know, so now you have synthetic influencers themselves.

Right.

So like these people don't even exist and they kind of look like they do and they're showing up in different places. And that was true, you know, last year images alone, but now it can be done with video and kind of be extended out like that. It's like, wow, what, you know, what I need for an influencer.

Yeah. I think in 2026 one of my predictions is going to be like we'll have the first AI actor like copyrighted or trademarked. Like somebody creates one of these fake influencers or something and it becomes so viral and popular that they actually like trademarked their AI. So like it's, oh, it's just like you have Mickey Mouse and these animations, right. That are copyrighted works the. I think you're going to have that with AI you're going to have like people will be monetizing the AI that they created and it's going to be unique to them and they'll sell it, license the likeness of it to other people to use.

You know, what do you think of the Disney and open AI getting together?

I mean it makes sense. I think it's great for Disney. They had to invest a billion dollars in open AI to make the deal happen.

It's weird. I thought the payment would come in the other direction.

Right.

You know, and it's like this is almost like New York Times saying, oh, we changed our mind. Take our content and let us, we'll pay you to take our content. You know, how did it flip around like that?

Well that, that just, that just proves the point. I mean, open it like, oh, the thing open AI has going for them right now and the chat GPT has going through it is that they own the consumer and they own the distribution. Right. And so people are spending time on this app and if you have content you want to get in front of million hundreds of millions of people, then you better be on the app.

And that's the thing that I left out at the beginning. We're talking about, oh, Anthropic and Google. Yeah. What OpenAI has going for it is that they have the platform. You don't go like, see the average person go say, you know, I'm going to get into Google AI Studio so that I can use. No, I'm going to go like, like get my credit card and get set up with an API key so, so I can do this. Comment. No, it's like, that's why in chat, GBT is like, we're gonna bring. Stay right where we are. We're gonna bring to you all of the stuff.

Right. Exactly. Now, that's just the power of platforms. And I do think chat has become a platform at this point. I'm comfortable saying that I think they've reached critical mass and they're gonna build it out and start to monetize it.

But why couldn't Google do that? I mean, this is what I said. Just want to get to, to picking on Google is that, you know, they've invented the transformer. They, they own this, the, the search engine world. They, they've got, they own YouTube. You know, they've got all this stuff. And yet why aren't they the platform? I go. People go to Google every day. Why didn't they go make it super easy for all this stuff to happen in a place where people don't have to think of something new. They just go and just do it.

I mean, I think one of the things Google has to navigate here is that, like, they were able to rely on the openness of the web to build a business model, and they never signed a license agreement with Disney or anybody else to show their stuff. Right. And everybody benefited. And they were able to monetize it with advertisements and eyeballs open AI. Like, I think Google is probably reluctant to start signing deals for content they would much rather follow. Like the YouTube model, where, hey, we'll give you the platform, the eyeballs, and then, you know, if you come and make content on our platform, we'll share some of the revenue with you. I mean, that I think would be the model that Google would rather see happen with AI.

Yeah.

That just explains why they wouldn't go do the Disney deal, which I heard came to them first. Or I don't know.

I don't know.

But what, you know, that figures. It doesn't fit their model to go and do that. But that doesn't explain why they're not the Destination, they're the destination for search. Everybody is like they're the only kind of site that doesn't have to worry about SEO because SEO is about getting to them, you know, and everybody and, and they're this. They're the kind of number one direct load site. People going google.com or going and using a browser where that's what that means anyway to be at that browser, put something in or put something in the URL bar at the top. It's like it means go to Google.

Right.

So they had the users, they had everybody coming straight to them and they didn't take advantage of it in this way.

Well, I think, I mean my pushback would be that they sort of have. Because search has re accelerated in the AI era. Like you know, user growth, paid clicks. People are using Google more now in the AI age despite the existence of chat GPT. And that is a big reason the stock has rallied because everyone sort of thought Google had missed this and search was screwed because of the chat bots. And it turned out that actually wasn't the case. Yeah, it could be a moment in time, I guess, but they, they seem.

To be doing right for a technology standpoint. Very well.

Right.

Very well indeed.

Cool. Do you guys have anything else? That was great conversation.

I did want to say, Bryce, to your point about, to your point about actors and stuff next year and Disney and all that, I think next year will be the first time that we hear a James Earl Jones's AI voice because he signed an agreement before he passed away to let Disney use his voice and an AI recreation of his voice. So I'm thinking next year we'll see a Star wars film with a pretty decent amount of lines that are AI generated and the uproar will be hilarious. We'll see how that works out.

Interesting.

John, last words. Anything you want to add?

Well, Scarlett Johansson was in a Marvel series of movies, I believe, and Disney owns Marvel and now OpenAI owns Disney. So the whole thing about not using her voice might come full circle.

And I mean that's just the thing. When you're an actor or actress and you sign up with one of these studios, you actually license over your likeness. Right.

I don't know. I, I mean I thought it was for that character, that kind of thing for sure.

But like Han Solo, like this. Harrison Ford doesn't own Han Solo. Disney does. Right. And now it like the, I think even in the open ideal they said there would be no voices. Right. And they were going to protect. And it's only Going to be like characters that don't have human faces or are animated characters.

Oh, okay.

I mean, come on. I think, I think Disney is fully within their legal right to license Han Solo, even his voice, even though it technically is Harrison Ford's voice. I mean, perspective. I think that's the case. Right. They can make a commercial. They can do. They have unlimited rights to do whatever they want with the Star wars franchise basically, without having to go back to the creator and be like, hey, would we, Harrison Ford? Can we use your voice as Han Solo? No, you've already agreed to that.

So, so last, last thing I'm thinking just is that I, maybe next time I can do more looking forward than just the agent thing. And. But this time looking backward and just to now and saying like 2025 was a year of incredible advances. If you go and think about where every modality you were talking about, you know, where it was in January 2025, where it is today, the kind of things that have happened, we've seen continual, really impressive growth in terms of capabilities throughout the year. If we keep seeing it like that, then we are on that curve that we talk about and we're just, you know, it's hard to see when you're on it, but like, we're already on it. We will look back and go, man, it's, it's been steep. You know.

It sure has. I totally agree with that.

Me too.

I think that's a wrap. That's a wrap on wrapped 2025. AI. So.

All right, well, I think so. Been good talking with you guys today and the whole year and looking forward to. To doing it in the new year.

Yeah. And Merry Christmas. Happy holidays to everybody out there, John, you included, and Craig, you know, Merry Christmas.

Merry Christmas. Yeah.

Take care, guys.

Bye.

Bye.